A recent AI in Testing Daily video on: Testing LLM and RAG Systems Evaluation, Golden Datasets, and Prompt Injection - Mar 10, 2026. This post highlights the core topic and why it matters, with the full discussion in the embedded video.
Why it matters: Tool-using agents turn untrusted text into a potential control channel. Even when models improve, the practical fix looks like classic security engineering: least privilege, compartmentalisation, and logs you can actually audit.
Singularity Soup Take: Prompt injection is basically social engineering for tool-using LLMs, and the long-term solution won’t be magic prompts—it’ll be boring, explicit security boundaries that assume the model will occasionally do something surprisingly dumb.
Watch on YouTube — AI in Testing Daily